If the model allows it, let's add few shot learning. see also: https://huggingface.co/openbmb/MiniCPM-V-2_6